Classificação e detecção de singularidades em imagens de impressão digital baseada em redes neurais convolucionais

• Ano de defesa: 2020
• Autor(a) principal: Silva, Paulo Ricardo Pereira da
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal da Paraíba; Brasil; Informática; Programa de Pós-Graduação em Informática; UFPB
• Palavras-chave em Português: Biometria; Impressão digital; Redes neurais convolucionais; Biometry; Fingerprint; Convolitional neural network; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: https://repositorio.ufpb.br/jspui/handle/123456789/18448

Biometry provides a reliable authentication mechanism using physical or behavioral traits to identify users based on their natural characteristics. Fingerprint recognition is one of the most used biometrics approach, since its high accuracy and low cost make the system more affordable and achieve satisfactory results. However, fingerprint recognition is still an open problem, since false acceptance and false rejection errors can be found in matching algorithm. This work proposes methods to classify and detect singularities from fingerprint images, that is based on Convolutional Neural Network and sliding window. The detection is done through sliding window that iterates through a input image and extracts a subimage of 50x50 and pass it as an argument to a classifier that labels between the classes loop, delta and neg or non-singularity, creating a set of candidate singularities and then filtering and keeping only true singularities - the classifier is a convolutional neural network and its architecture is similar to LeNet, with the first convolutional layers and the last are fully connected. To evaluate the performance of the proposed method and the enhancement effect over images, 2 groups of database were used: the first with no pre-processing that is formed of the databases FVC2000-3, FVC2002-1, FVC2004- 1, FVC2006-3, and SPD2010; the second one with pre-processing that is formed of the databases FVC2000-3, FVC2002-1, FVC2004-1 and FVC2006-3. The results showed the enhancement did not affect on significantly the performance in detection. Over the group with no pre-processing, the best classifier reached the value 99% of accuracy and 0.98 of average F-score for three classes, and the detector reached 0.8 of F-score and median IoU of 0.73. And the best classifier over the group with pre-processing reached the value 99% of accuracy and 0.99 of average F-score for three classes, and the detector reached 0.84 of F-score and median IoU of 0.76.